MAPIE's sample_weight vs sklearn's sample_weight in the fit method

[CarlaFernandez]

How do these parameters relate?
If my original model relied on sample_weight, can I simply wrap MAPIE around it and it would be the same internally?
Like this:

self.mapie_regressor = MapieRegressor(estimator=self.model, method='plus', cv=5, random_state=self.seed)
        
self.mapie_regressor.fit(
    X=self.X_train, 
    y=self.y_train, 
    sample_weight=self.weights
)

[thibaultcordier]

Hello @CarlaFernandez,

TL;DR: These parameters are the same for MAPIE and scikit-learn in fit method.

Indeed, MapieRegressor or other MAPIE classes have been implemented to be scikit-learn compatible with commonly used terms. In particular, sample_weight is used in MAPIE specifically in the utility function fit_estimator that depends on your model estimator.

In other words, if you simply wrap MAPIE around your original model which relies on sample_weight, it will be the same internally (direct transmission of sample_weight parameter).

Some additionnal links:

• https://scikit-learn.org/stable/glossary.html#term-sample_weight
• https://scikit-learn.org/stable/modules/generated/sklearn.base.RegressorMixin.html
• https://scikit-learn.org/stable/modules/generated/sklearn.base.ClassifierMixin.html
• https://stackoverflow.com/questions/54093185/how-to-understand-sample-weight-in-sklearn-metrics
• https://github.com/scikit-learn-contrib/MAPIE/blob/master/mapie/utils.py#L79C3-L79C3

[CarlaFernandez]

Thanks for the confirmation!